Use case with measurements: CONFIG_NUMA_EMU=y gives 30-60% more
throughput on a memory-bound Pi 5 workload, and on 7.0 it is a config
change only

  Answering the question Juerg Haefliger asked on 2026-09-10 (comment 14
on this bug).

  We run an SDR signal-processing workload on a Pi 5 8GB under Ubuntu 26.04: 
numpy FFTs over 60 s blocks of radio samples, several processes at once. It is 
  memory-bandwidth bound: one process alone scales fine, but concurrent 
processes contend.
  
  We rebuilt linux-raspi from tag Ubuntu-raspi-7.0.0-1023.23 through the Ubuntu 
packaging, with one annotation override:
  
    CONFIG_NUMA_EMU    policy<{'arm64': 'y'}>
  
  No patch was needed: on 7.0, NUMA_EMU depends on NUMA_MEMBLKS and 
GENERIC_ARCH_NUMA, which the raspi config already sets. The firmware's 
numa=fake=8 then 
  produces 8 nodes, and numa_policy=interleave applies.
    
  Same hardware and same workload, two runs on each kernel, alternated stock, 
emulation, stock, emulation. Whole-machine throughput (work units per minute, 
  median of three rounds, higher is better):
  
    processes   stock 7.0.0-1023   with CONFIG_NUMA_EMU=y
    at once     run 1 / run 2      run 1 / run 2
    1           2.14 / 2.04        2.26 / 2.24
    2           2.49 / 2.22        3.18 / 3.08
    3           2.11 / 2.01        3.10 / 3.06
    4           1.84 / 1.65        2.77 / 2.88
  
  Every emulation round at 2+ processes beat every stock round. A single 
process is essentially unchanged; the gain is in contention between processes. 
The 
  kernel booted through piboot-try validation and ran about an hour of this 
benchmark without failures.
    
  As far as we can tell, the annotation that disables it on arm64 comes from 
LP: #1864198, a 2020 s390x request that gives no reason and names no other 
  architecture. Raspberry Pi's own kernel enables it for the Pi 5 
(CONFIG_NUMA_EMU=y in bcm2712_defconfig).
  
  Would you consider enabling CONFIG_NUMA_EMU for linux-raspi arm64?

-- 
You received this bug notification because you are a member of Ubuntu
Bugs, which is subscribed to Ubuntu.
https://bugs.launchpad.net/bugs/2166421

Title:
  Malformed early option 'numa' passed in numa=fake=8

To manage notifications about this bug go to:
https://bugs.launchpad.net/ubuntu/+source/linux-raspi/+bug/2166421/+subscriptions


-- 
ubuntu-bugs mailing list
[email protected]
https://lists.ubuntu.com/mailman/listinfo/ubuntu-bugs

Reply via email to